c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 7 ( 2 0 1 4 ) 218–224

j o ur na l ho me pag e: www.int l .e lsev ierhea l th .com/ journa ls /cmpb

A smart medication recommendation model forthe electronic prescription

Shabbir Syed-Abdula, Alex Nguyena, Frank Huanga, Wen-Shan Jianb,Usman Iqbala, Vivian Yangc, Min-Huei Hsua,∗, Yu-Chuan Lia,∗

a Taipei Medical University, College of Medical Science and Technology, Graduate Institute of Biomedical Informatics,Taiwanb Taipei Medical University, School of Health Care Administration, Taiwanc College of Medical Science and Technology, Institute of Biomedical Informatics, Taiwan

a r t i c l e i n f o

Article history:

Received 13 January 2014

Received in revised form 4 June 2014

Accepted 27 June 2014

Keywords:

NHI database

Medications

Inappropriate prescription

Diagnosis-Medication association

Smart medication recommendation

model

a b s t r a c t

Background: The report from the Institute of Medicine, To Err Is Human: Building a Safer

Health System in 1999 drew a special attention towards preventable medical errors and

patient safety. The American Reinvestment and Recovery Act of 2009 and federal criteria of

‘Meaningful use’ stage 1 mandated e-prescribing to be used by eligible providers in order

to access Medicaid and Medicare incentive payments. Inappropriate prescribing has been

identified as a preventable cause of at least 20% of drug-related adverse events. A few studies

reported system-related errors and have offered targeted recommendations on improving

and enhancing e-prescribing system.

Objective: This study aims to enhance efficiency of the e-prescribing system by shortening

the medication list, reducing the risk of inappropriate selection of medication, as well as in

reducing the prescribing time of physicians.

Method: 103.48 million prescriptions from Taiwan’s national health insurance claim data

were used to compute Diagnosis-Medication association. Furthermore, 100,000 prescrip-

tions were randomly selected to develop a smart medication recommendation model by

using association rules of data mining.

Results and conclusion: The important contribution of this model is to introduce a new concept

called Mean Prescription Rank (MPR) of prescriptions and Coverage Rate (CR) of prescrip-

tions. A proactive medication list (PML) was computed using MPR and CR. With this model

the medication drop-down menu is significantly shortened, thereby reducing medication

selection errors and prescription times. The physicians will still select relevant medications

even in the case of inappropriate (unintentional) selection.

© 2014 Elsevier Ireland Ltd. All rights reserved.

∗ Corresponding authors.E-mail addresses: drshabbir786@gmail.com (S. Syed-Abdul), alexn

jj@tmu.edu.tw (W.-S. Jian), usman.iqbal85@gmail.com (U. Iqbal), vjaak88@gmail.com, jack@tmu.edu.tw (Y.-C. Li) .http://dx.doi.org/10.1016/j.cmpb.2014.06.0190169-2607/© 2014 Elsevier Ireland Ltd. All rights reserved.

thhp@gmail.com (A. Nguyen), frank0202@gmail.com (F. Huang),yang@tmu.edu.tw (V. Yang), 701056@gmail.com (M.-H. Hsu),

i n b i o m e d i c i n e 1 1 7 ( 2 0 1 4 ) 218–224 219

1

Etefmveordatamuawa

hhadwpsp

hpiosdIcftp

esism

2

Tctomiwn

c o m p u t e r m e t h o d s a n d p r o g r a m s

. Introduction

lectronic Prescription (ePrescription, eRx) is one of the func-ionality/module of the computer-based order entry (CPOE) orlectronic health record system (EHR) used by health providersor prescribing, transferring, dispensing and monitoring of the

edications [1]. Barach et al., reports that nearly 100,000 indi-iduals die per year in the U.S. due to preventable medicalrrors [2], and medication errors were found to be a major typef medical error [3]. The release of the Institute of Medicineeport, To Err Is Human: Building a Safer Health System in 1999rew a special attention towards preventable medical errorsnd patient safety [4]. The advancement in the informationechnologies, particularly health information exchange withinnd among hospitals has been proven to significantly reduceedical errors and improve patient safety [5]. A decade later,

nder the American Reinvestment and Recovery Act of 2009nd federal criteria of ‘Meaningful use’ stage 1, e-prescribingas mandated to be used by eligible providers in order to

ccess Medicaid and Medicare incentive payments [6,7].Reduced medication errors and improved patient safety

ave become the important indicators used for evaluatingospital performance and for approval of ‘meaningful use’ccreditations. Studies have reported more than an 80% ofecrease in the medication errors in the in-patient settingsith the use of ePrescription system [5,8,9]. Therefore, e-rescribing has been promoted as a potential informationystem in order to reduce medication errors and increaseatient safety [8,10].

Although on the one hand there are studies published thatave evaluated the ePrescription system and reported theositive results [5,11,12], on the other hand there are stud-

es revealing CPOE systems possess potential risk for 22 typesf medication error [13] Up to 35% of prescribing errors wereystem-related (selection of an inappropriate (unintentional)rug from the drop-down menu next to a likely drug)[14].

nappropriate prescribing has been identified as a preventableause of at least 20% of drug-related adverse events [15,16]. Aew studies reported system-related errors and have offeredargeted recommendations on improving and enhancing e-rescribing system [17,18].

Therefore, this study is aimed to enhance efficiency of the-prescribing system by reducing the risk of inappropriateelection of the medication and also to reduce the prescrib-ng time of the physicians. In order to do so we proposes amart model that recommends most commonly prescribededications in the drop-down menu for a given diseases.

. Method

aiwan’s national health insurance claim data was used toompute the Disease-Medication association for all prescrip-ions from 1st January to 31st December 2002. Out of a totalf 263.57 million prescriptions from out-patient clinics 160.09

illion were excluded for the following reasons: (a) miss-

ng/invalid disease code or medication code; (b) medicationritten in the Mandarin; and (c) prescription of traditional Chi-ese medications. About 103.48 million prescriptions with the

Fig. 1 – Study design.

diagnosis in a valid (International Classification of Disease v.9– Clinical Modification) ICD9-CM code and medications in thenational health insurance (NHI) codes were analysed. Taiwan’sNHI medication codes were mapped into ATC (AnatomicalTherapeutic Chemical) Classification System in order to quan-tify disease-medication association. Fig. 1 shows the studydesign. The NHI permits up to three diagnoses, thereforeusually a prescription consists of 1–3 diagnoses, and 1–15 med-ications.

A smart medication recommendation model was devel-oped by the following steps:

1. To quantify the disease-medication (DM) association fre-quencies from the 103.48 million prescriptions.

2. To calculate the DM ‘interestingness’ (lift) value and thenranked value for each of the 14,734 diseases.

3. To compute the Mean Prescription Rank (MPR) and Cover-age Rate (CR) of the medications from randomly selected100,000 prescriptions.

4. To decide Coverage Rate of medications by alphabeticalorder for the proactive medication list (PML).

5. To demonstrate the user interface of the model.

2.1. Step 1: To quantify the disease-medication (DM)association frequencies

All the 103.4 million valid prescriptions were selected in orderto compute frequencies of the associations for every diagnosiswith all the medications prescribed. Each prescription consistsof 1–3 ICD9-CM codes and 1–15 medications. The 103.4 millionprescriptions had 14,734 unique ICD9-CM codes for diseaseswith 3657 ATC codes for medications resulting in 53.8 millionunique associations and a total of 798.5 million associations.

2.2. Step 2: To calculate the DM ‘interestingness’ (lift)value and ranked for each of the 14,734 diseases

Association rule mining was performed on structured claim

database of the NHI. The popular association measure of inter-estingness of a rule ‘lift’ was used [19,20]. A lift value greaterthan 1.0 indicated that a prescription with diagnosis A tended

220 c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 7 ( 2 0 1 4 ) 218–224

Table 1 – An example of a prescription with three diagnoses and five medications with calculated ‘lift’ value ranks for DMassociation.

Medication ICD9-CM

461.9 462 466.0

Ordinary salt combinations and antiflatulents M1 165 189 220Cephalexin M2 76 85 125Mefenamic acid M3 105 95 161Opium derivatives and expectorants M4 99 83 70carbinoxamine M5 20 47 38

Mean Prescription Rank (MPR) = AVG (165, 76, 95, 70, 20) = 85.2.Bold indicates minimum ranks of the medication for the given pair of diagnosis in this particular prescription.

1),

to contain medication B more often than prescriptions that didnot contain A. The lift values for all the DM association werecalculated and sorted in descending order for each disease,thus the most commonly used medications for every diseasewere obtained. Then, all the lift values were ranked (DM valueby descending order) for each disease.

2.3. Step 3: To compute the Mean Prescription Rank(MPR) and Coverage Rate of the medications

100,000 prescriptions were randomly selected for further anal-ysis. In this step analysis was done on each prescription withall of the DM associations present in it. We have introduceda new term called Mean Prescription Rank (MPR). MPR is theaverage value of the lowest DM rank of lift value of each pre-scription (See Table 1). Since prescription could include up tothree different diagnoses, it is necessary to computed lift valueof each medication used for various combinations of diseases.In order to know the most frequently prescribed medicationsfor a given combination of diagnosis. MPR was calculated forall of the 100,000 prescriptions. Next, Coverage Rate of pre-scriptions was computed from 100,000 MPRs computed.

MeanPrescription Rank = AVG (MIN (Lift-rank (ATCi, ICD9CM

where Lift-rank is the function to get rank of Lift-value and ithe number of ATC codes.

Coverage Rate (CR): When each of the 100,000 prescrip-tions was computed using MPR, it was found that the top100 MPR prescriptions covered up to 81% the total 100,000prescriptions, and the top 200 MPR covered up to 96% ofthe total prescriptions. This suggests all the medications pre-scribed in top 100 MPR will be included when we computeproactive medication list. Table 2 shows the CR of prescrip-tions in relation to the MPR.

Fig. 2 shows the graphical representation of Table 2. It

shows that the top 50 MPR are covered by 45% of prescrip-tions, the top 100 MPR are covered by 81% of prescriptions andthe top 200 by 96% of prescriptions.

Table 2 – Mean Prescription Rank (MPR) to the CoverageRate for the 100,000 prescriptions.

MPR 22 42.5 64.5 100 129 200 846.5Coverage Rate 20% 40% 60% 81% 90% 96% 100%

Lift-rank(ATCi, ICD9CM2), Lift-rank (ATCi, ICD9CM3))i)

Fig. 2 – Graphical representation of MPR and Coverage Rateof the medications in the prescriptions.

2.4. Step 4: To decide Coverage Rate of medications byalphabetical order for the proactive medication list (PML)

The complete medication list contains the names of all themedications prescribed in the 100,000 prescriptions and sorted

by alphabetical order. However, proactive medication list iscomputed from the medication present in the prescriptionsof top 100 or 200 MPR. For an example, if the hospital man-agement decided to have the medications prescribed in top100 MPR prescriptions for their ePrescription system, thenmedication prescribed in these top 100 MPR will be sortedin alphabetical order and presented in the drop-down menuof the ePrescription system. The hospital management willknow that if they choose the top 100 MPR, medication pre-scribed in 81% of prescriptions will be covered; however, ifmanagement choose the 200 MPR medication prescribed in96% of prescriptions will be covered. Table 3 shows the num-ber of medications by alphabetical order presented in thecomplete medication list (CML) and the pro-active medica-tion list (PML) for the top 100 and 200 MPR prescriptions andthe percentage of medication list shrinkage by each letter,respectively.

As shown in Table 3, the total number of medicationsstarting with the letter ‘A/a’ are 268; however, with theuse of PML, on an average there will be 12 medications

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 7 ( 2 0 1 4 ) 218–224 221

Table 3 – Display number of medications by alphabetical order for complete medication list (CML), proactive medicationlist (PML) for the top 100 and 200 MPR prescriptions and their percentage of shrinkage, respectively.

Char CML PMLTop 100 MPR

Shrink100 (%) PMLTop 200 MPR

Shrink200 (%)

a 268 11.63 95.66 20.61 92.31b 193 10.42 94.60 17.72 90.82c 414 20.46 95.06 36.90 91.09d 258 10.43 95.96 17.85 93.08e 164 6.21 96.21 11.05 93.26f 182 8.37 95.40 15.10 91.70g 95 8.24 91.33 11.15 88.26h 94 3.92 95.83 6.90 92.66i 164 7.95 95.15 12.25 92.53j 1 0 100 0 100k 15 1.58 89.49 2.61 82.60l 112 7.31 93.47 11.83 89.44m 316 13.04 95.87 23.52 92.56n 116 8.35 92.80 12.96 88.83o 104 4.43 95.74 7.98 92.33p 378 12.3 96.75 22.69 94.00q 19 1.39 92.69 2.04 89.27r 105 5.93 94.36 9.60 90.86s 215 8.18 96.19 14.37 93.32t 324 11.08 96.58 19.37 94.02u 11 1.38 87.49 2.02 81.63v 56 3.32 94.07 4.59 91.81w 1 0.93 7.20 0.98 1.80x 12 1.07 91.05 1.26 89.51y 5 0 100 0 100

92

ip

2m

Ttptaarem

3

TmdevTtttmw

z 35 2.51

n the top 100 MPR and 21 medications for top 200 MPRrescriptions.

.5. Step 5: To demonstrate the user interface of theodel

he user interface with the complete and proactive medica-ion list (PML) is shown in Fig. 3. With the use of PML, thehysician can select medications from a list of 21 medicationshat are commonly prescribed and not from a drop-menu with

complete list of 268 medications. Even in cases of inappropri-te (unintentional) selection, the physician will still select theelevant medication. Thus, this model will reduce medicationrror and the prescription time by significantly shortening theedication drop-down menu.

. Discussion

his study proposes a smart medication recommendationodel for electronic prescriptions. This model organizes the

rop-down menu of the ePrescription system. It lists only rel-vant medication for a given diseases and also displays a liftalue to show the relation of the medication to a given disease.his model is smart because it not only reduces physicians

ime for prescribing medications by significantly shortening

he drop-down menu but also eliminates the risk of medica-ion errors even if Physicians unintentionally select the wrong

edication. This is because only most relevant medicationsill be listed in the menu. This model is developed to be a

.84 3.49 90.03

solution for the inefficiency of the present ePrescription sys-tems, such as providing possibility of inappropriate selectionof medications [14–18].

The associations rule mining technique we used is not anovel by itself. It has been developed for over a decade andhas been used in a various fields and purposes [20–22]. Itwas used to infer patient problems from clinical and billingdata, to estimating co-occurrence of the disease and find-ings from discharge summaries, to discovering potentiallyunknown adverse effects of medications as well as for pub-lic health surveillance [23–26]. However, none of the studieshave reported the use of association rule mining techniquefor finding co-occurrence of diseases and medications as weused in order to develop a model to recommend medicationlist relevant to the diagnoses.

The important contribution of this model is to introduce anew concept called Mean Prescription Rank of prescriptions. Itcould have been a simple solution to compute most frequentlyused medication for each diagnosis; however in practice theprescription may have up to three or more different diagnoses.Therefore, Mean Prescription Rank was computed in order toovercome the medication errors for the prescriptions with amultiple diagnoses. MPR provides hospital management a cal-culated risk of exclusion or inclusion percentage of the totalmedications used by the hospital or clinic. To the best of ourknowledge none of the available studies to date have proposed

any model to deal with medication errors in a prescriptionwith a multiple diagnosis.

In case a physician is willing to prescribe a new or themedication that is not listed in the PML, the system will

222 c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 7 ( 2 0 1 4 ) 218–224

Fig. 3 – A screen shot of the drop-down menu with the complete and pro-active medication lists for three prescriptions witha single diagnosis.

i n b i

acr

4

TnaiiIdipscvoimta

C

Aesfiani

E

N

F

N

A

Sfj

r

prediction models using feature analysis, Acad. Radiol. 12

c o m p u t e r m e t h o d s a n d p r o g r a m s

utomatically switch to the CML. This happens once the physi-ian starts typing the letters not matching with medicationsecommended in PML.

. Limitations

his model is based on analysis of only one year data fromational health insurance claim data in Taiwan. With thedvancement in the drug innovation, inclusion of new drugsnto prescriptions is a common clinical practice. Therefore, its recommended to re-compute the DM lift values annually.n this study MPR and PML was computed from 100,000 ran-omly selected prescriptions. It would be interesting to know

f 100 million prescriptions are taken for analysis would haverolonged the PML. Generalizing the DM lift values for pre-criptions done out of Taiwanese hospitals is also a matter ofoncern. However, this model can be useful to analyse DM liftalues and compute MPR in order to organize PML. As a partf our future work, we want to enhance this model by strat-

fying medications prescribed according to sex and age. Thisodel could be more robust and comprehensive once medica-

ions are integrated with the appropriate dose and frequencyccording to sex and age of the patient.

ompeting interest

ll authors have completed the Unified Competing Inter-st form at www.icmje.org/coi disclosure.pdf and declare: noupport from any organization for the submitted work; nonancial relationships with any organizations that might haven interest in the submitted work in the previous three years,o other relationships or activities that could appear to have

nfluenced the submitted work.

thical approval

ot required.

unding

one.

ppendix A. Supplementary data

upplementary data associated with this article can beound, in the online version, at http://dx.doi.org/10.1016/.cmpb.2014.06.019.

e f e r e n c e s

[1] R.A. Miller, R.M. Gardner, K.B. Johnson, G. Hripcsak, Clinical

decision support and electronic prescribing systems: a timefor responsible thought and action, J. Am. Med. Inform.Assoc. 12 (2005) 365–376.

o m e d i c i n e 1 1 7 ( 2 0 1 4 ) 218–224 223

[2] P. Barach, S.D. Small, Reporting and preventing medicalmishaps: lessons from non-medical near miss reportingsystems, BMJ 320 (2000) 759–763.

[3] E. Ammenwerth, P. Schnell-Inderst, C. Machan, U. Siebert,The effect of electronic prescribing on medication errorsand adverse drug events: a systematic review, J. Am. Med.Inform. Assoc. 15 (2008) 585–600.

[4] T.A. Brennan, The Institute of Medicine report on medicalerrors – could it do harm? N. Engl. J. Med. 342 (2000)1123–1125.

[5] D.W. Bates, et al., Reducing the frequency of errors inmedicine using information technology, J. Am. Med. Inform.Assoc. 8 (2001) 299–308.

[6] Minnesota Department of Health & Office of HealthInformation Technology, http://www.health.state.mn.us/e-health/eprescribing/erx032011guidance.pdf, 2011.

[7] T.K. Gandhi, et al., Outpatient prescribing errors and theimpact of computerized prescribing, J. Gen. Intern. Med. 20(2005) 837–841.

[8] A.M. Gilligan, et al., Analysis of pharmacists’ interventionson electronic versus traditional prescriptions in 2community pharmacies, Res. Soc. Adm. Pharm. 1–10 (2012).

[9] D.W. Bates, et al., Effect of computerized physician orderentry and a team intervention on prevention of seriousmedication errors, JAMA 280 (1998) 1311–1316.

[10] N. Staggers, C. Weir, S. Phansalkar, in: R.G. Hughes (Ed.),Patient Safety and Quality: An Evidence-Based Handbook forNurses, Agency for Healthcare Research and Quality,Rockville, MD, 2008.

[11] E. Oren, E.R. Shaffer, B.J. Guglielmo, Impact of emergingtechnologies on medication errors and adverse drug events,Am. J. Health Syst. Pharm. 60 (2003) 1447–1458.

[12] A.L. Potts, E.E. Barr, D.F. Gregory, L. Wright, N.R. Patel,Computerized physician order entry and medication errorsin a pediatric critical care unit, Pediatrics 113 (2004)59–63.

[13] R. Koppel, et al., Role of computerized physician order entrysystems in facilitating medication errors, JAMA 293 (2005)1197–1203.

[14] J.I. Westbrook, et al., Effects of two commercial electronicprescribing systems on prescribing error rates in hospitalin-patients: a before and after study, PLoS Med. 9 (2012)e1001164, http://dx.doi.org/10.1371/journal.pmed.1001164.

[15] D.W. Bates, et al., Incidence of adverse drug events andpotential adverse drug events, JAMA 274 (1995) 29–34.

[16] R. Tamblyn, et al., The medical office of the 21st century(MOXXI): effectiveness of computerized decision-makingsupport in reducing inappropriate prescribing in primarycare, Can. Med. Assoc. J. 169 (2003) 549–556.

[17] K.C. Nanji, et al., Errors associated with outpatientcomputerized prescribing systems, J. Am. Med. Inform.Assoc. 18 (2011) 767–773.

[18] T.S. Lesar, L. Briceland, D.S. Stein, Factors related to errors inmedication prescribing, JAMA 277 (1997) 312–317.

[19] PMML 3.0 – Association Rules. DATA Mining Group,http://www.dmg.org/v3-0/AssociationRules.html (accessed17.07.12).

[20] A. Wright, E.S. Chen, F.L. Maloney, An automated techniquefor identifying associations between medications, laboratoryresults and problems, J. Biomed. Inform. 43 (2010) 891–901.

[21] J. Iskander, V. Pool, W. Zhou, R. English-Bullard, V. Team,Data mining in the US using the vaccine adverse eventreporting system, Drug Saf. 29 (2006) 375–384.

[22] J.A. Carrino, L. Ohno-Machado, Development of radiology

(2005) 415–421.[23] A. Wright, et al., A method and knowledge base for

automated inference of patient problems from structured

s i n

328–337.[26] I.M. Mullins, et al., Data mining and clinical data

224 c o m p u t e r m e t h o d s a n d p r o g r a m

data in an electronic medical record, J. Am. Med. Inform.Assoc. 18 (2011) 859–867.

[24] H. Cao, M. Markatou, G.B. Melton, M.F. Chiang, G. Hripcsak,Mining a clinical data warehouse to discover disease-finding

associations using co-occurrence statistics, AMIA Annu.Symp. Proc. (2005) 106–110.

[25] X. Wang, G. Hripcsak, M. Markatou, C. Friedman, Activecomputerized pharmacovigilance using natural language

b i o m e d i c i n e 1 1 7 ( 2 0 1 4 ) 218–224

processing, statistics, and electronic health records: afeasibility study, J. Am. Med. Inform. Assoc. 16 (2009)

repositories: insights from a 667,000 patient data set,Comput. Biol. Med. 36 (2006) 1351–1377.